G. Sudhakar Babu, Dr. K. Chandra Sekhara Reddy / International Journal of Engineering Research and Applications (IJERA) ISSN: 2248-9622 www.ijera.com Vol. 3, Issue 4, Jul-Aug 2013, pp.870-876
Automatic License Plate Recognition System G. Sudhakar Babu1, Dr. K. Chandra Sekhara Reddy2 Abstract: Automatic license plate recognition system (ALPR) is one of the most important of the intelligent transportation system (ITS). A number of techniques have been used for vehicle plate character recognition. The proposed system uses image processing and neural network character recognition and pattern matching of character as two character recognition techniques. In this approach multilayer feed forward back propagation algorithm is used. The performance of the proposed algorithm has been tested on several vehicle plates and provides very satisfactory results. Keywords: Automatic license plate recognition system (ALPR), intelligent transportation system (ITS). I.
Vehicle license plate recognition (VLPR) is an image processing system whereby it is used to recognize the vehicles by identifying the license plate. It is basically used for traffic and security purposes. The cycle will start when the vehicle steps over the detector. It will activate a signal to the vehicle license plate system of the presence of the vehicle. The illumination will be activated and images of the front picture of the vehicle will be taken. The system will read the information pixels of the vehicle and run the recognition process and system will apply error backpropagation algorithm to analyze the vehicle image. Besides analyzing, the images will be enhanced, locating the vehicle plate position and extract the characters from the vehicle plate. The characters will be recognized by using neural network. Then system will try to match the recognized vehicle plate number with the car plate database. If access granted, the gantry will open and allowed the vehicle. Previously different neural models were designed to filter the noisy sign. So, many researches of car identification have been approached by car license plate extracting and recognition, some of the related work is as follows. Lotufo et. al  proposed automatic number-plate recognition using optical character recognition techniques. Johnson and Bird  proposed Knowledgeguided boundary following and template matching for automatic vehicle identification. Fahmy  proposed bidirectional associative memories (BAM) neural network for number plate reading. Itâ€™s appropriate for small numbers of patterns. Nijhuis et. al  proposed fuzzy logic and neural networks for car LPR. This method used fuzzy logic for segmentation and discretetime cellular neural networks (DTCNNâ€™S) for feature extraction. Choi  and Kim  proposed the method based on vertical edge using Hough transform (HT) for extracting the license plate. E.R.Lee et. al  used
neural network for color extraction and a template matching to recognize characters. S.K. Kim  used a genetic algorithm based segmentation to extract the plate region. Tavsanoglu et. al  proposed an approach to form orientation map as recognition feature using a Gabor filter for recognizing characters. Yoshimura et. al  used Gabor jets projection to form a feature vector for recognizing low resolution gray-scale character. Hontani et. al.  proposed a method for extracting characters without prior knowledge of their position and size in the image. Park et. al.  devised a method to extract Korean license plate depending on the color of the plate. H.J. Kim et. al  proposed a method of extracting plate region based on colour image segmentation. In this study, the proposed approach is based on extraction of plate region, segmentation of plate characters and recognition of characters.
The license plate recognition process can be roughly divided into three steps as shown in Figure1, Plate Localization, Character Segmentation and Character Recognition. Each step will be carried out by an independent module. An input image submitted to the system is first examined and processed to obtain the vehicle license plate region, then the plate region is process to locate each individual digit and character, these are then submitted to the final Optical Character Recognition (OCR) process to determine the identification.
Figure1: Recognition steps III.
This is the process of extracting the license plate region out of an image. It involves basic image processing techniques combined with some decision making based on deterministic threshold. Without any prior knowledge on how large the plate is, or where it is located, the entire image must be inspected and analyzed in order to extract candidate
870 | P a g e
G. Sudhakar Babu, Dr. K. Chandra Sekhara Reddy / International Journal of Engineering Research and Applications (IJERA) ISSN: 2248-9622 www.ijera.com Vol. 3, Issue 4, Jul-Aug 2013, pp.870-876 regions. There are many different approaches on how to accomplish this, some algorithm assume that the plate region location of the image should not varies by much , and it should be adjusted by using sensors , thus limiting the search range for fast results. Some techniques make use of only edge information for plate location , and there are also very complex algorithms such as vector quantization , fuzzy clustering  and fuzzy logic.
Character Segmentation Once the candidate region is detected from the input image, the next stage is to segment the plate to extract individual digit/character for recognition. This process is highly dependent on the format of the plate being processed. Because different country and regions have different plate shapes and sizes, the color used as plate background and foreground are totally different and their content varies both in length and combination of digit and characters. For example, the Chinese license plates analyzed in  have dark background with character in lighter color, the algorithm used cannot be directed applied to the plates in Alberta, Canada , since those are completely opposite. Some techniques do exist for specific cases and can be adapted to other cases, such as combination of vertical and horizontal projection to determined glyph location , or by using an adaptive clustering technique by finding white spaces between columns of higher density of dark pixels . However, these techniques do not take into account that sometimes there could be frames that are partially connected to the plate content. Take as an example Figure 2, which features a plate that has a frame that is connected to the lower part of the middle five glyphs. The use of simple projection or clustering techniques will not yield adequate results.
Figure 2: Framed plate
Character Recognition Once each individual digit and/or character is extracted, it must be identified in some way. This process is called Optical Character Recognition, and there are several different solution to this problem. Two approaches are particularly popular among many different researches on license plate recognition. One of the methods is template matching . In this method, a series of slightly different templates of all glyphs are kept in a database. Once a image is submitted for recognition, existing templates are compared to the new image,
and the best fit will determine its identity. This method requires that the template database be large enough to cover most glyph variations, and it must also have an efficient algorithm to process large set of templates. The other way is by using Artificial Neural Networks as classifier . The ANN classifier must be trained before use to recognize all the different digits, characters and symbols that require identification. Several other approaches are also being used, feature analysis is the method used in , which is an analysis of the characters to distinguish it. And for recognizing a totally different set of symbols in those plates in Thailand, Hausdorff Distance technique was the choice .
System Architecture: Each country and region has its own specifications for their legal vehicle license plates, and they differ from each other in several aspects. To begin with, the license plate itself can have different shape/size, have totally opposite set of colors for background and foreground  and it can be located at different location on the vehicle. Also, the number of digits or characters will differ, A region with only a couple of thousands cars only need to use a plate with 6 or 7 numeric digit, but for a region with several million cars that configuration is not nearly enough. This can change the location of each digit/character on the plate, they can be in a specific font, on some specific background colors or not. For a country that does not use Latin characters as their primary written language, even totally different set or symbols are needed . These substantial differences among them make it almost impossible to create a single algorithm that can be universally used for all plates. The algorithms defined in this chapter are designed to recognize California State license plates as shown in Figure 7.3. The regular California license plate has the following configuration The plate background is a single uniform color with no graphics. The word “California” appears at the top-center on the plate in red and italic. The plate number starts with a digit (0-9), followed by three English characters (A-Z) and three more digits (0-9). All seven digit/characters are sans type font. The plate background is a light shade of gray while its characters are of dark bluish color.
Figure 3: California License plate
871 | P a g e
G. Sudhakar Babu, Dr. K. Chandra Sekhara Reddy / International Journal of Engineering Research and Applications (IJERA) ISSN: 2248-9622 www.ijera.com Vol. 3, Issue 4, Jul-Aug 2013, pp.870-876 Plate Localization
The algorithm proposed does not make any assumption regarding possible location and/or size of the plate region (See Appendix A for some sample images of different cars). It only relies on both edge and color information for extracting the license plate region out of random images. It makes use of some constants that provide helpful information on the plate location in any image (Fig 4). No matter the lighting condition, plate background colors are all lighter and plate foregrounds are darker. Plate region are all rectangular shape with same specific proportion between width and height. Plate region will be high in edge concentration. The following a flow diagram of the algorithm used (Figure 5)
Images can be taken at various lighting conditions. In each case, the difference in brightness will greatly influence the color pixel value of the plate and the characters. This process is an attempt to compensate for the different lighting conditions under which the image is taking (Fig 7.6 & 7.7). This is achieved by applying the following: 1. Convert the image into HSV color space 2. Equalize its intensity component. 3. Convert back into RGB color space
Figure. 6 Intensity Histogram before Equalization Figure.7 Intensity Histogram after Equalization
Extract Edge Information Figure 4: Sample input image
None of the traditional edge detection operations such as derivative based detection or canny edge detection is used. Instead, the edges are revealed by creating a difference image img as where src is the grayscaled source image, threshold is a binary threshold function and close(src) is the morphological close operation. This process enhances high frequency features in the image, thus highlighting the license plate content. This image provides a good starting point on where to locate the plate (Fig 8).
Figure.5 Plate location
Noise Filtering Images acquired through a digital camera may be contaminated by a variety of noise sources. In image processing terms, noise is refer to stochastic variations as opposed to deterministic distortions such as shading. While most current cameras will reduce the noise levels to a very low level, noise can never be eliminated completely. Noise can be further reduced by application of the smoothing filter. In this case, a 3x3 Gaussian convolution filter with σ = 0.5 is used.
The next step is to filter the image eliminating backgrounds and unnecessary elements. This is accomplished by a combination of using the edges information obtained from previous step and some assumption about California license plates. 1. From previous image, the glyphs in the plate are clearly presented. And next to these pixels, the plate background can also be extracted. 2. As mentioned, California state license plates has a light background which will be captured as different shades of gray depending on lighting conditions. A color is picked according to above two conditions, and it is assumed to be the plate background color and used as filter (Fig 9).
Connected Component Analysis These patches are converted into binary images, and grouped as connected components. Components of different size and shapes will be scatted around the image at this point and one of
872 | P a g e
G. Sudhakar Babu, Dr. K. Chandra Sekhara Reddy / International Journal of Engineering Research and Applications (IJERA) ISSN: 2248-9622 www.ijera.com Vol. 3, Issue 4, Jul-Aug 2013, pp.870-876 them contains the license plate being searched. The analysis process will find the candidate region out of the collection of connected components by using some previous knowledge about the plate. A component is considered to be a candidate if 1. The area covered by it also covers an area of high concentration of non-zero pixels from Figure 8, an indication of possible presence of many character/digits. 2. And its width/height proportion is similar to the CA license plate. Any connected component matching these two criteria is considered to be a candidate region worthy of further examination. This process can produce results ranging from zero to several candidate regions. If many candidate regions are found, they are all considered to be valid candidates and submitted to the next stage of processing. As shown in Figure10, out of all the different connected components created on the filtered image, two of are identified as candidates. Both are considered valid regions and pass along to the character segmentation stage. If no candidate region is found, the process is repeated from 3.1.4 by picking a different color and using the new “plate background” color for filtering. New connected components of different size and shape will be obtained and analyzed. If the system still fails to find any region after multiple colors has been used, the image is rejected by the system as lacking a plate region.
valid set of character/digit regions are found in one of them.
Figure 10 Connected component analysis
Apply Color Filter The candidate region is filtered by eliminating the colors that do correspond to neither the background nor foreground. This filter is applied again to eliminate the red “California” text at the top of the plate. It also eliminates other foreign objects on the plate, such as stickers.
Figure.11 Character segmentation Figure.8 Image Revealing Edges
Thresholding Use an adaptive algorithm to find a threshold for this image and use a inverse thresholding function to create a binary image.
Figure 9 Filtered image
In the best case scenario, the only pixels left after the binarization would be the pixels forming the License number (Figure12).
Character Segmentation The segmentation algorithm is a combination of several techniques that are applied as necessary. First, a simple method like adaptive binary thresholding is used along with some clustering. If results are negative, then some other methods are added progressively increasing its complexity. The high level flowchart of the algorithm is shown in Figure 11. The following steps are repeated for each possible candidate region until
Figure 12 Filtered plate
873 | P a g e
G. Sudhakar Babu, Dr. K. Chandra Sekhara Reddy / International Journal of Engineering Research and Applications (IJERA) ISSN: 2248-9622 www.ijera.com Vol. 3, Issue 4, Jul-Aug 2013, pp.870-876 Characters Extraction Create connected component out of the filtered image and use boxes to fit the components. Try to extract the glyphs by finding a series of boxes with very similar size and shape. For a successful image, seven boxes should be found (Figure 7.13), each one will contain a group of pixels forming a digit or a character. If the number of boxes found is less than seven, then one of two possible routes is selected. 1. If more than two boxes have been found, then an interpolation is performed to reconstruct the required seven boxes based on existing boxes. 2. If less than two boxes are found within the plate region and there is a large cluster of pixels, then this could indicate the presence of a frame attached to the plate content in some ways. An attempt for frame removal is performed, and the image is re-evaluated starting from 3.1.1. 3. If above steps are applied several times without success, the candidate region is discarded and the next region from the collection is evaluated. After individual characters are extracted, each character is converted into 20x10 matrices of one's and zeros. This matrix will be submitted to the final stage of recognition by neural network.
Figure14 Multi-layer Feed-forward network
Back-Propagation Algorithm The term is an abbreviation for "backwards propagation of errors". It is one of the most used methods of supervised training for ANN. It is most useful for training multilayer feed-forward neural networks . The algorithm can be summarized as follows 1. Introduce input training data to the neural network. 2. Compare the network's output to the desired output from that input. Calculate the error in each output neuron by
3. Calculate the error at each weight and adjust 4. Propagate error values to upper levels
Simulated Annealing Algorithm
Figure 13 Character extraction
Character Recognition The Optical Character Recognition method chosen as part of the LPRS system is an Artificial Neural Network. NEURAL NETWORK ARCHITECTURE: ANN can be divided into categories based on their different topologies Single-layer neural network: neuron receives input and process to produce output. Multi-layer Network (Figure 14): One input layer, one output layer with or without hidden layer(s). This architecture is also called Feed forward as information flow is strictly one way. Each neuron receives signal from layer above and fire signal to layer below. Self-Organized: Network connections are dynamically established or broken to produce the desired outcome.
Similar to the metallurgical process, each step of a Simulated Annealing algorithm replaces the current solution by a random "nearby" solution, chosen with a probability that depends on a global parameter T (called the temperature). T is gradually decreased during each step of the process. Initially, when T is high, the system will basically change almost randomly around the entire search space. But as T decreases, the system will come into an configuration which is the best of all the configurations tested. The random wandering saves the system from becoming stuck at local minima. Thus the temperature decrease takes long enough, the system will eventually come down to a globally optimum solution. Neural Network Classifier Implementation Two different neural networks are used in LPRS for recognizing the glyphs extracted in the segmentation stage. One is used only to recognize numbers, and the other is trained to recognize exclusively characters. The choice of separate neural networks was made specifically to target California license plates, which has predetermined placement for digits and characters. Separated neural network means that each network have to learn to recognize less symbols, this expedited the training process and reduced errors rate. Both neural network are Feedforward networks, and use Sigmoid function to
874 | P a g e
G. Sudhakar Babu, Dr. K. Chandra Sekhara Reddy / International Journal of Engineering Research and Applications (IJERA) ISSN: 2248-9622 www.ijera.com Vol. 3, Issue 4, Jul-Aug 2013, pp.870-876 control neuron firing. The character neural network has one input layer, one hidden layer and one output layer, the input layer is a 20x10 matrix of neurons, the hidden layer has 5 neurons, finally the output layer size is 26, each representing a character in the English alphabet. The digit neural network also has three layers, and has the same input layer size. Its hidden layer has 4 neurons and its output layer size is 10, one of each digit. The training on the neural networks is carried out by using a combination of classical error back-propagation and simulated annealing process. Each Training block in the diagram is a shortened back-propagation training process in which the ANN will gradually and steadily converge “downhill”. After the ANN has received some training, it is submitted to the annealing process to avoid local minima. Some random neighboring samples are chosen by altering neuron weights within a normal range based on global temperature T. The Energy of each neural network is obtained by calculating the overall error (σ) as Where Ei is the results of the error function evaluated for the ith training data set and rand(T) is a random value based on T. By adjusting the sum of error, “uphill” movement is made possible. The network with the lowest energy (lowest σ) is considered to be the best configuration at that point and will received more back-propagation training. The process is repeated until T reaches zero or σ reaches zero. The following is the flow chart of the ANN training algorithm implemented. Simulated annealing are proven to have the ability of not getting stuck of any local minima, but the random nature of its sampling heuristic does not always guarantee that system is moving toward best possible direction. In the other hand, backpropagation almost guarantees a fine granular descent into a nearly best, but that nearby best is not always the global optimal. By combining the two methods, this training heuristic produces better and faster results than using either method separately, the system will be altered to the placed around a configuration near the globally best solution, and quickly descend into it.
Experimental Results: The computer used for running the experiment is a notebook, with a Pentium 4 1.0 GHz processor and 512 MB of RAM. Due to the limited availability of data at the time of experiment, the LPRS system was tested by using a reduced set of data. A total of 50 images with vehicle license plates taken under different lighting conditions, different distances, and different exposures were used as input to the system. The images were originally taken using a 2 Mega-pixel digital camera and resized to 800x600 for quicker processing.
Plate Localization All images were submitted to the plate localization process, candidate regions are marked with a red rectangle after processing (See Appendix A for sample image). The images are visually inspected to verify its correctness (Table 1). Total Plate Failed to Success Images located locate rate 50 48 2 96% Table 1: Plate Localization Success Rate The reason of two of the images failed was exclusively because of the quality of the image (See Appendix B for sample image). Once picture was blurred and the other picture was taken from a distance and as a result, the plate area was very small. Character Segmentation Only those images with candidate regions identified are submitted to Character Segmentation. After processing, the plate area is extracted from the input image and all seven boxes enclosing the characters are drawn in red (Table 2) Total images
Cumulative success rate
48 45 3 93.75% 90% Table 2: Character Segmentation Success Rate Recognition rate on training data 80% Recognition on actual images extracted 60% from LPRS Table 3: OCR Recognition Rate IV.
The proposed method for the vehicle number plate recognition using the back propagation approach, showed the remarkable enhancement in the performance when two hidden layers were used. The recognition accuracy is best in experiments where MLP with two hidden layers was used. Number of hidden layers is proportional to the number of epochs. This means that as the number of hidden layers is increased; the training process of the network slows down because of the increase in the number of epochs. If the accuracy of the results is a critical factor for an vehicle number plate recognition application, then the network having many hidden layers should be used but if training time is a critical factor then the network having single hidden layer should be used. The proposed approach could be used with conjunction with other ones for better security and increasing the area of the particular application.
875 | P a g e
G. Sudhakar Babu, Dr. K. Chandra Sekhara Reddy / International Journal of Engineering Research and Applications (IJERA) ISSN: 2248-9622 www.ijera.com Vol. 3, Issue 4, Jul-Aug 2013, pp.870-876
R. Plamondon and S. N. Srihari, ''On-line and off-line handwritten character recognition: A comprehensive survey'', IEEE. Transactions on Pattern Analysis and Machine Intelligence, vol.22, no. 1, pp. 6384, 2000.  R. A. Lotufo, A. D. Morgan, and A. S. Johnson, 1990, “Automatic Number-Plate Recognition,” Proceedings of the IEE Colloquium on image analysis for Transport Applications, vol. 1, no. 35,pp.6/1-6/6, February 16, 1990.  A. S. Johnson, B. M. Bird, 1990, “Numberplate Matching for Automatic Vehicle Identification” IEE Colloquium on Electronic Image and Image Processing in Security and Forensic, April, 1990.  M. M. M. Fahmy, 1994, “Automatic Number-plate Recognition: Neural Network Approach,” Proceedings of VNIS’94 Vehicle Navigation and Information System Conference, 3 1 Aug-2 Sept, 1994.  J. A. G. Nijhuis, M. H. Ter Brugge, K. A. Helmholt, J. P. W. Pluim, L. Spaanenburg, R. S. Venema, M. A. Westenberg,1995, “Car License Plate Recognition with Neural Networks and Fuzzy Logic,” IEEE International Conference on Neural Networks, 1995.  H. J. Choi, 1987, “A Study on the Extraction and Recognition of a Car Number Plate by Image Processing,” Journal of the Korea Institute of Telematics and Electronics, Vo1.24, pp. 309-315,1987.  H. S. Kim, et al., 1991, “Recognition of a Car Number Plate by a Neural Network,” Proceedings of the Korea Information Science Society Fall Conference, Vol. 18, pp. 259-262, 1991.  E. R. Lee, P. K. Kim, and H. J. Kim, 1994, “Automatic Recognition of a Car License Plate Using Color Image Processing,” Proceedings of the International Conference on Image Processing.  S. K. Kim, D. W. Kim, and H. J. Kim, 1996, “A Recognition of Vehicle License Plate Using a Genetic Algorithm Based Segmentation,” Proceedings of 3rd IEEE International Conference on Image Processing, V01.2., pp. 661-664, 1996.  V. Tavsanoglu, E. Saatci “Feature extraction for character recognition using Gabor-type Filters implemented by cellular neural networks,” Proceedings of the 2000 6th
G.Sudhakar Babu, Assistant Professor Dept of Physics, SVIT, Hampapuram, Anantapur. He has 11 years of teaching experience He received his M.Phil from Allagappa University, Tamilnadu, He published 1 papers in International Journal. His areas of interest are Embedded systems and Semiconductors.
Dr. K. Chandra Sekhara Reddy is working as Associate Professor of Physics in S.S.B.N.College, Anantapur. He is having 25 years of teaching experience. He received his Ph.D. from Sri Krishnadevaraya University, Anantapur in the year 1996. He guided 11 M.Phil students. Five scholars are working under his guidance for their Ph.D. He published 9 papers in National and International Journals. His areas of interest are Liquid Crystals, Nano materials and Embedded systems.
876 | P a g e
Published on Jul 6, 2013